Método Automático de Detecção de Nódulos Pulmonares em Imagens de Tomografia Computadorizada do Tórax Usando Redes Neurais Convulocionais

Abstract

Cancer is the second leading cause of death in the world. Among all types, lung cancer remains the most frequent and the leading cause of death due to cancer, whose greater difficulty in treatment is related to the stage of diagnosis. In many cases, the disease is identified at an advanced stage, when current treatments are not effective. Despite initiatives to promote early diagnosis, physicians do not always make the best use of information obtained from medical imaging devices. Limitations of the human visual system, insufficient training and experience, factors such as fatigue and distraction may contribute to the inefficient use of available information.

In this scenario, automatic techniques for the analysis and processing of medical images can be applied as a diagnostic tool. The main goal of this thesis is to propose an automatic method to detect pulmonary nodules from thorax computed tomography images using convolutional neural networks. Initially, optimal thresholding, connected component analysis, and mathematical morphology operations are used to isolate the lungs from the whole image. The lungs are then analyzed to identify the higher density internal regions. Samples of pulmonary nodules and healthy regions are extracted from axial, sagittal, and coronal examination planes and concatenated to compose an orthogonal image used for training of a convolutional neural network.

Subsequently, a set of forty features extracted from nodule candidates is used for false-positive reduction in an SVM pattern classifier. The main contributions of this thesis are the proposal of a validation system composed of four sets of gold standards created from the LIDC-IDRI database, the creation of an automatic method for reducing the area of interest from the lung segmentation, the proposal of a convolutional neural network topology for training and detection of pulmonary nodules, the proposal of a set of characteristics extracted from nodule candidates to reduce false positives, and finally, pulmonary nodule detection and segmentation.

The method's performance is evaluated using the 10-fold cross-validation technique at LUNA16 and LIDC-IDRI databases, obtaining a sensitivity of 84.56% and 74.04–81.54%, respectively. The results allow us to conclude that convolutional neural networks are effective in the detection of pulmonary nodules.